video_OCP2026_和碩_數位孿生到量產就緒_20261015

Hello. Hello. Hi. Good morning. Morning. I'm John from Pegatron product planning. Uh today I'm going to share some something we are going to deliver in this year. Uh we use the digital twin in AI factory infrastructure to improve production readiness. Uh in short is we build a day and decide a virtual platform which can help us to optimize our factory readiness. The The key point is we want to enable a fast deployment for a new factory and lower the risk and the cost, of course. All right. Uh before diving for the more detail, probably many folks here is not really familiar. I am a giant from Taiwan. Uh quick briefing is Pegatron is a global global engineering and manufacturing company. We serve the global leading brands including electronic consumer electronics, communication device like a mobile phone and the AI server we are building now. So, uh in global we have 21 global sites around the world. So, that is our current challenge. We you know, building a factory is extremely expensive. We cannot waste any resource and time on this. So, we decide to build a virtual platform and to help us to improve the speed to speed out the factory deployment. So, to address this, we have been thinking about what if everything would digitalize in a virtual space first. And then we then we can identify the end-to-end system bottleneck at the very early stage. So, we call the the left side of the slide we call the Pegatron AI factory. And next, we apply the design to real factory on the side and right side. Uh then we send the data back to the simulation. So, you can see it's like a closed loop. The virtual side on the left and we'll do the simulation and design and the apply the design on the new factory on the right side and new factory they will calculate on the age age computing system at factory local side and send the data back to the virtual virtual platform. So it's like a closed loop solution. We evaluate, design, apply and send the data back to the system and keep optimized. It's like closed loop. All right. How does it work of AI factory of the digital twin synchronization? On the left side of first and second layer you can see the first is a digital twin service we provide a model. That's what I just mentioned. We decide we design development virtual space. A second is AI service transfer data into the intelligence for the AI agent development. A third one is a smart factory service connect the data for the operation connect data connect connect the operation data from the physical factory. As I said we use the factory have the sensor. They have the age computing collect data. We will summarize the data and send it back to our virtual space. So on the right is the AI power physical factory operation operation data flows back to the platform then we collect and we analyze re-analysis and keep improving. Okay, so All right. Uh because we have built from six virtual factory already in different region. A benefit of that is our teams because our teams can in different region they can plan, simulate and collaborate in the same digital platform. In one same digital platform. That's very important because if you have factory in India India they want to build a factory a brand new factory they can leverage the experience and the data the factory in the China. So they can waste no time and develop it in very soon. All right. So let's take a look at our Let's take an example of our Taoyuan factory. That's the place we build our AI rack server. Uh when we build a factory, you can see the product uh button that pick uh we bought a brand new factory. It's purely empty without any facility and equipment inside inside. And then in in the middle, we simulate in the virtual space, we design and apply to the most right picture to become a brand new factory. So, in overall is the exact data we just experienced, uh we can reduce the construction time around 40%. So, this video can help you to more understand how to how we are doing now. Uh supposed there is a should be a sound. Left side is the cyber space, and right side is about a physical uh factory. You can see it's really really similar. Let's pick cuz uh that's because we get we build a model exactly same as our device. So, the model the facility, the testing equipment dimension, height, and width, everything is exactly same. That can help us to deploy and apply on a production line very very accurate. The distance between the equipment will be very accurate to make sure everything's correct. All right. Um it's now the production line uh we're talking about the area or something. Uh this is another example I want to share with you is we we model the power chain for the facility. What's that? It's thermal lab. Because you know, the rack the rack level power consumption is extremely high. If you have a one bunch cluster, uh you have the liquid cooling device, you have high or cool IO, everything you need to consider you need to consider in your thermal lab. So, it's you can also have your engineer to do the setting to consider liquid cooling condition about the CDU operation, and liquid flow, liquid pressure, and temperature. So, best the best scenario is you can do scenario you can do simulate in different IT load. Probably 50%, 70%, or 100% full loading to ensure your simulate is designed on right condition. All right. Uh we have another tool uh the right side of the page is a web UI. Uh our engineer our engineer can use this web UI to adjust all the setting right here about the CDU liquid cooling and pressure and voltage and under the different workload to make to see the final result. And system will alert you if the current uh similar labs similar result is out of the device spec. It can help you to measure everything you design in the new new lab it meet your expectation. All right, quick recap uh the component of the solution I just mentioned. The one is uh air factory air factory the virtual space we enable we enable the AI agent to learn. The second is digital twins enable agent to practice. And the real world I just mentioned about the fist the real factory uh we will provide a feedback for continuous improvement. So, machine learning, machine practice and we everything done in the virtual world and apply on the physical factory. So, uh the virtual space I just I just talked about is the Pixel own solution, but we can also to partner with an external vendor to provide more AI solution. This is one example I want to share with you uh the cable routing. Because one class three if you have 16 uh variable being racks, how many cables are there? Couple of hundreds, right? And we have we partner with our external cable external vendor to provide the auto routing. You will see the picture you will see the video later and we have a robo robo arm that can automatic pick out the cable and plug in the right direction right right pole and connect to another rack. So the first first thing is about Let me pull that up. Let me see. Uh digital twin The first component we have three components right here. The first is a digital twin. The second is we leverage an external vendor Newton physical engine. Uh it provide a safety environment to do the AI testing. And next will be the advanced robotic AI. It's about a visual language action model uh performing interaction environment variety. So how does it work? You can see the bottom up it is workflow. That's how we design this system. Uh we first you have to get a CAD and define the requirements back. Next we will auto generate you know auto generate a cable routing path to tell you how the path how the cable how long distance of the cable should be and how to how to connect from this rack to another rack. And next we will send the data into the our virtual space to do the digital twin simulation. Next will be the motion planning and simulate before implementation. And finally we will send to the uh uh validation test to the physical robot. And you can see the arrow circle back to the motion planning. That's that's the same scenario I just mentioned. Uh when every time we validate on the physical machine every arrow every issue they will send the data back to uh to re-simulate again to find the best answer. That's why maybe you can see the video later. First time the cable is not plugged to the right pole they probably blocked by the other cable. They will make adjustment and insert again successfully. >> Open improvement and adaptive intelligence. >> Pro- is our project. Project Jeda is our project name. Is not virtual Is not virtual video. Is exactly exactly the same our factory. >> We address this. Peg-A-Verse Jeda utilizes high-fidelity physics to build policies within a virtual environment. By capturing human expert demonstrations, imitation learning serves as the foundation for policy bootstrapping, distilling core task logic into a reliable baseline for autonomous operations. Through massive parallel reinforcement learning in NVIDIA Isaac Lab, Jeda iterates through vast numbers of scenarios to master the dynamics of flexible materials. Residual learning bridges the sim-to-real gap, neutralizing physical variances to ensure a stable transition to the real-world environment. Powered by NVIDIA Isaac Root N Open VLA model, Jeda integrates visual and multimodal force observations. With integrated force sensing, the system detects real-time resistance and provides tactile feedback when obstacles or mis- >> This is about just mention. You always sometimes you will block by other cable. You will You will have a sense of the real the and readjustment of direction to plug to the right cable or plug to the right hole. >> Jeda synchronizes dual arm policies with VLA to bridge the sim-to-real gap through an autonomous observation thinking action. Jeda agent redefines the industrial frontier, evolving automation into autonomy. Peg-A-Verse Jeda Think before you act. >> All right. Um just want to use a real video to show you is not a just theory. It's not just a theory. It's a thing we are actually doing in our factory. So, this is the last page I want to wrap up. Um I just share we use our digital platform to create a virtual space. Uh we apply our real factory and we apply our desired new thermal lab. And now that is auto cable by the robot arm. So, Pegatron is almost 20 years 20 years old and in 20 years ago is a long production line, 50 operators to do the manual assembly notebook desktop server. And now we are doing the machine learning and machine practicing. Everything is going to robot arm. Maybe in the nearly future I think our factory uh all the factory run over will be lights out and produce the AI server 24/7. So, anyway, our target is to enable fast deployment, reduce the implementation, and continuous requirement uh to improve our factory. Our last Thank you. And our booth is a year two. If any further question you want to talk about neural server solution, we can talk about later. Thank you.